Latent fingerprint software improvements save police time with higher accuracy

Latent fingerprints are the original raison d’etre for the biometrics industry, but the technology behind their collection and matching is subject to the same forces as other modalities and use cases. A whole new class of software has been submitted for the Evaluation of Latent Fingerprint Technologies (ELFT) by the U.S. National Institute of Standards and Technology this year, reaching best-yet performance in multiple aspects.

Idemia Public Security Deputy CTO and Head of Strategic Innovation & IP Vincent Bouatou tells Biometric Update in an interview that the test measures “The ability of the technology, mostly algorithms, to take that fingerprint mark, which is a very low quality, very imperfect image of a fingerprint, and match it against a database of significant size.”

The process evaluated by ELFT involves fully automatically ingesting an image, extracting relevant information and comparing it to the reference database to find a match, he explains.

This automated process is ultimately in aid of the certified latent print examiner, who conducts the manual examination to make a final decision on whether a match is found, and whether the confidence with which that match is assessed is sufficient to qualify as evidence.

Examiners historically spend a long time, looking until they find a match, or determine that there is no match to find, Bouatou says.

The improvement seen in the ELFT suggest that a law enforcement officer can feed a fingerprint directly into the system, without supplementary markup, and the system will return likely candidates. “It will be far better at making sure that the right candidate, if the candidate is in the database, is in the first ranks, so that the number of verifications that the latent fingerprint examiner will have to do will be very low.”

Accuracy is not the only measure of latent fingerprint software’s effectiveness, and even judgements about accuracy require some nuance. “Quote-unquote” accuracy usually quoted is rank-one accuracy, according to Bouatou.

Another aspect of accuracy is the ability of the system to say, “don’t bother checking, there’s nobody that close in the record”

Crime scene investigators often collect hundreds of latent prints, Bouatou says, but “maybe only a few of them will actually reveal evidence that is useful for the investigation.”

“If you can just automatically process all of these latents, and just look at the ones where there is potentially a determination to be made, it is a ton of time that is being saved for the latent examiner.”

The number of prints the expert examiner will consider depends on several factors, Bouatou says, including the severity of the crime being investigated and the quality of the latents collected.

“If you have a very high-quality mark, then you expect the system to be able to bring the candidate within rank one or two,” he says. “So, if you don’t find the guy in rank one, rank two, maybe three, you will move onto the next search, because in all likelihood he wasn’t in the database.”

With a very low-quality mark the system will struggle, forcing the examiner to go further down the list. Bringing the right candidate up the list, therefore, saves considerable time.

Training and testing for the real world

Idemia trains its latent fingerprint processing algorithms with quality as representative of what will be found in the field as possible.

Bouatou emphasizes the importance of confidence to the process. An examiner sometimes thinks they have matched a latent to the right person, but will say the match is indeterminant because they can’t say with the confidence necessary to meet the threshold for evidence that it is. In this case, a match at a high rank does not necessarily help.

Latent fingerprint software development is very different from training biometric software for something like a border control system, Bouatou says. It is almost as different as a completely separate modality.

For one important difference, “We tend to throw a lot more computing power at latent identification than we do at regular fingerprint identification.”

Bouatou draws an analogy with the moon landing for a comparison between latent and live fingerprint matching.

“To solve a very particular problem you had to invent technology that didn’t exist, that you wouldn’t necessarily have invented otherwise,” he explains. “Once you have, you can apply it to a lot of other use cases.”

“Latent fingerprint is sort of the moonshot of fingerprint identification, and there’s a lot of technology trickling down to the most mundane use cases we also address within Idemia.”

Bouatou has spent 25 years with Idemia – roughly half as long as the company has been working on latent fingerprint matching software.

In that time, the introduction of modern AI techniques just within the last 10 years stands out as a disruptive event. Idemia introduced AI to fingerprint biometrics cautiously “because we wanted to make sure that we have a level of control over the technology that we had over the more classical algorithms, if you will,” Bouatou says.

The benefits of AI in fingerprinting are only now being realized, as opposed to in facial recognition where they were immediate.

Those benefits are particularly needed in latent fingerprint processing. In contrast with other modalities, which Bouatou notes have reached “ninety-nine point something percent” accuracy in independent testing, “latent work is still very close to our hearts because it’s the hardest.”

Future developments should bring better assessment of the probability of a hit prior to sending it to the matching engine, telling CSIs, “Don’t bother sending this to the AFIS, save the electricity,” Bouatou says.

He also expects further improvements to the processing of very low-quality prints. In those situations, even if a match is not enough for a determination on its own (i.e. as evidence), it could help lead to other information that is. Idemia also remains focused on keeping the footprint of system small.

Synthetic data is not particularly useful for training latent fingerprint software, or other biometric software, Bouatou says. It is useful, however, to augment data for testing.

Bouatou lauds Idemia’s work on algorithmic bias reduction, and NIST’s efforts to test for it.

This is one area where synthetic data is thought by some to hold some promise, but Bouatou says Idemia’s work in more than 100 countries provides experience and opportunities for training data on diverse populations. This is also important for sensor development, he points out.

The company uses its own employees consented data a lot, according to Bouatou. They represent about 80 nationalities.

In testing, synthetic data can be used to control the parameters of what’s being generated “very precisely.” Idemia also provides technologies in development to partners, in some cases, so they can test their system for features like usability. Partners can customers can also use their own data, which Idemia does not have access to, for testing, freeing Idemia from concerns around the handling of personally identifiable information and regulatory compliance.

Even after testing, Bouatou says some innovation takes place in partnership with partners and customers, such as to improve usability. This is the genesis of Idemia’s mobile fingerprinting capturing technology.

“We’re really trying to invent what’s next with the practitioners and the end-users,” Bouatou says.

Article Topics

biometric matching  |  biometric testing  |  biometrics  |  ELFT  |  fingerprint biometrics  |  forensics  |  IDEMIA  |  NIST